Open julian-xinchi opened 3 years ago
Highly unlikely for v1.0.
The original vision for RISCV vector did propose matrix multiplication, including outer product. However, the Vector TG decided that vision was too visionary. The result is a comprehensive but limited set of augmented 32bit instructions. Direct matrix operations will be a later release, planned to use 64bit instruction encoding.
V1.0 does provide features that support efficient matrix multiplication operation code sequences. Notably these include:
Highly unlikely for v1.0.
The original vision for RISCV vector did propose matrix multiplication, including outer product. However, the Vector TG decided that vision was too visionary. The result is a comprehensive but limited set of augmented 32bit instructions. Direct matrix operations will be a later release, planned to use 64bit instruction encoding.
V1.0 does provide features that support efficient matrix multiplication operation code sequences. Notably these include:
- stride load/store to efficiently load the columns of arrays.
- shift and move under mask to duplicate the row data by the number of columns. -- whole register move [copy] between physical vectors registers for use when LMUL>1
- For full matrix [not just outer product] there are [3 argument] fused multiply add variants.
Sweet! Thanks for your reply. Maybe i more expect the version after v1.0 :-)
https://en.wikipedia.org/wiki/Outer_product